JAIST and Princeton researchers developed the NTAC algorithm to classify neurons by their connections

Researchers from JAIST, Princeton and other centers presented NTAC — a system that determines neuron type not by external form, but by its synaptic connections. In experiments on fruit fly connectomes, the algorithm showed accuracy above 90% and completed the task in minutes on a regular laptop.

Why form is misleading

Neuron typing has long been stuck on manual work. Usually experts classify cells by morphology: shape, branching, position in tissue. This doesn't always work.

In some brain areas, especially in the visual lobe of the fruit fly, different neuron types look almost identical. It's hard to distinguish them visually, although functionally they participate in different signal processing chains. Because of this, annotation becomes a long and expensive process that doesn't scale well as connectomes grow.

The authors suggest looking not at the cell itself, but at its "connections." The logic is simple: two identical wires in a wall are hard to distinguish by appearance, but easy to tell apart if you trace where they go. Neurons have a similar story.

NTAC uses the synaptic connection pattern as the main feature and shows that it better reflects the functional nature of the cell than its anatomical silhouette. This is especially important in dense brain regions with repeating architecture.

How NTAC works

NTAC operates in two modes. In the semi-automatic variant, researchers pre-label a small fraction of neurons, and the model transfers this knowledge to other cells in the same dataset. In the second mode, annotation isn't needed at all: the algorithm clusters neurons itself based on connection similarity. This matters for large connectomes where manually labeling all cells is already impossible. This approach allows the use of both partially ready atlases and completely raw datasets.

• Semi-automatic mode uses some ready-made labels
• Unsupervised mode clusters neurons without hints
• Runtime takes minutes on a standard laptop
• The method was tested on several fruit fly connectomes

According to the authors, the model was tested on FlyWire datasets, the visual lobe connectome, and ventral nerve cord data. This is not a toy example on a single lab sample, but a comparison on several modern maps of the fruit fly nervous system. Such a design makes the result stronger: researchers demonstrate not a one-time success, but an approach capable of transferring between different areas, different tasks, and different numbers of cell types in the sample.

Results and limits

NTAC showed its strongest results where morphological methods most often stumble. In the visual lobe, the semi-automatic variant exceeded 90% accuracy, whereas the popular NBLAST approach, which relies on neuron shape, hovered around 50%. In fully unsupervised mode, NTAC achieved about 70% accuracy in complex areas, whereas morphological clustering in some cases stayed below 10%. For automatic annotation tasks, this is a very notable gap.

"The connection pattern itself carries sufficient signal for rapid

neuron type identification."

It's also important that this isn't a resource-hungry model for a data center. The authors emphasize that NTAC runs on standard CPUs and doesn't require a supercomputer. For neurobiology this is a practical shift: if connectomes grow from fruit fly brain to mouse and beyond to human, automation of cell typing will become mandatory. The algorithm was already used to label thousands of neurons, and the researchers name mouse brain mapping as the next major milestone.

What this means

NTAC doesn't solve the problem of mapping the entire human brain, but eliminates one of the slowest stages — manual cell typing. If the approach maintains accuracy on larger organisms, connectomics will get a working tool that will accelerate both basic science and the search for neural circuit disruptions in diseases.